acceptodds
Under review as a conference paper at ICLR 2027

Wrong Rules, Right Predictions: Concept Remapping and Repair in Neuro-Symbolic Models

Abstract

Neuro-symbolic models combine neural perception with symbolic knowledge, but what happens when the symbolic knowledge itself is wrong? A model can compensate for incorrect rules by changing the meaning of its learned concepts while preserving task performance. On MNIST addition, permuted rules retain 97.5% label accuracy even as joint concept accuracy falls to 35.8%. We study what can be certified about such misspecified knowledge before training and whether these guarantees translate into recovery during learning. Given a reference for the intended semantics, we certify whether concept remapping can compensate for an incorrect program and whether the intended program is its unique minimum-edit repair. For full-support addition with concepts, we prove that the nearest competing zero-loss program differs in exactly table entries, yielding a tight unique-repair guarantee for up to corruptions. Yet identifiability does not imply recoverability. On ten previously unseen identifiable twenty-edit tables, a residual encoder recovers the intended program in 48/50 MNIST runs but only 10/50 SVHN runs, while a smaller CNN recovers 46/50 on both. Failures often coincide with the collapse of a single-digit concept. These results separate structural compatibility, repair identifiability, and neural recoverability under misspecified symbolic knowledge.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Neuro-symbolic learning, concept remapping, repair identifiability, rule misspecification, neuro-symbolic reasoning, minimum-edit repair, concept learning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_tLJ7OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes to study neuroSymbolic models whose supplied background knowledge KK (a table mapping concept pairs to labels) differs from the intended K⋆K^{\star}, while leaving training labels intact. The proposed investigation spans three layers: compatibility, repair identifiability and neural recovery. Compatibility asks whether there is a concept remapping reconciling the supplied knowledge KK with the true knowledge K⋆K^{\star}. Repair identifiability refers to the possibility of finding the unique repair of the corrupted knowledge at minimum cost and, finally, neural recovery to whether the training actually recovers this repair. The authors define a certified edit margin Δ\Delta, prove that for 2-slot addition over nn digits the nearest competing zero-loss table is 2(n−1)2(n-1) edits distant from K∗K^*, and report MNIST/SVHN experiments showing that the repair is not always recovered by end-to-end training.

Critical strengths

The problem the paper addresses is interesting per se and understudied, relative to the reasoning-shortcut literature that assumes K=K∗K = K^*. The results in Table 2 are clear in this sense: they show how, under the permutation corruption of the knowledge, the label accuracy stays high while concept recovery is severely affected. The characterization of the minimum-repair in Theorem 3 is correct and tight (for the case of a sum between two digits). The paper also comes with an extensive reproducibility package.

Critical weaknesses

W1: Theorem 3 is the paper’s only non-trivial theoretical result and it already fails to carry over to the comparison task.

The guarantee it provides is addition (2 digits) specific; already for the case of comparison in Proposition 6 a single corrupted cell can create ties. So outside of addition, the framework provides no guarantee and everything falls back to CP-SAT that solves using K∗K^* anyway (the ground truth).

W2. Certificates have limited practical value.

They provide a quantitative measure of how different KK and K⋆K^{\star} are, but they need to be computed using K⋆K^{\star}, which is not available in practice. Given Eq.(5), stating that Δ>0\Delta > 0 implies K∗K^* is the unique minimum-edit explanation follows by definition; the only substantive content is computing ealte_{alt}.

W3. Theory is narrow and trivially follows from the assumptions.

Theorem 3 is the main theoretical result, but it is very narrow in scope: it only covers the case of addition between two digits. There is no result beyond two inputs, n-ary programs and logic programs as DeepProbLog. Many results, like Proposition 1 and 2, are elementary derived from the assumptions and do not carry significant insight.

W4. Contribution 3 is not supported by the experimental design.

The main failure mode is attributable to the encoder choice, as Table 3 shows that SlotCNN recovers the intended knowledge 46/50 times on both MNIST and SVHN, while a SmallResNet succeeds only 10/50 times on SVHN. Then the surrogate recovery is measured only in terms of the objective in Eq. (6) (apart from an ABL adaptation in App E.5), and this objective is also not properly motivated nor justified. So general conclusions about the recovery under neural training are not supported by the experiments.

W5. Experiments are narrow both in the scope of the data (simple MNIST and SVHN datasets) and the knowledge (2 digits addition, comparison).

The paper mentions [1], and yet never considers more challenging tasks from that benchmark suite.

W6. Severe problems with the writing and presentation.

  • ΔΔ, e∗e* and ealte_{alt} are used at (071–077) before ed(⋅,⋅)ed(·,·) is defined (214–219).
  • Several assertions are made without either a justification or a reference. As an example, consider the Jensen argument (178) and lines (233-234). These are just two of several others I found throughout the paper.
  • Proposition 1 bundles together four different statements at once.
  • The experimental section simply reads the data tables and provides little discussion about the meaning of the results.

W7. Related work is neither complete nor accurate.

Semantic Loss is cited as a reasoning shortcut mitigation, but it is not. Conversely, several mitigation strategies from [2] are not reported. Also the paragraph does not include a relevant method for addressing the issue in ABL as [3].

The paper picks an interesting problem, but the weaknesses in theoretical scope, applicability, experimental design, writing and presentation are major. I recommend rejection.

References

[1] Bortolotti, S., Marconato, E., Carraro, T., Morettin, P., van Krieken, E., Vergari, A., Teso, S., and Passerini, A. (2024). "A neuro-symbolic benchmark suite for concept quality and reasoning shortcuts." Advances in Neural Information Processing Systems 37 (Datasets and Benchmarks Track).

[2] Marconato, E., Teso, S., Vergari, A., and Passerini, A. (2023). "Not all neuro-symbolic concepts are created equal: Analysis and mitigation of reasoning shortcuts." Advances in Neural Information Processing Systems 36.

[3] Huang, Y.-X., Dai, W.-Z., Cai, L.-W., Muggleton, S. H., and Jiang, Y. (2021). "Fast abductive learning by similarity-based consistency optimization." Advances in Neural Information Processing Systems 34.

Questions for authors

Have you considered extending the framework to more than 2 concept inputs? In particular, does the permutation argument underpinning Theorem 3 still hold for k > 2 slots?

Reviewer_8UnwOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper considers a special case for neuro-symbolic models: the encoded symbolic knowledge is wrong. They study what can be certified about such misspecified knowledge before training and whether these guarantees translate into recovery during learning.

Critical strengths

They investigate a scenario that is common in practice for neuro-symbolic models: the encoded symbolic rules are wrong.

Critical weaknesses

The overall writing needs to be improved. Many descriptions are ambiguous; thus, the contributions remain unclear. The performance is examined on very limited datasets.

Questions for authors
  1. The overall writing needs to be improved. There are many ambiguous statements in the paper. When reading the introduction section, I find it hard to understand these parts: (1) L50: What does “remapping” mean here? Do you mean retraining or fine-tuning the concept predictor? (2) L51: What is the actual meaning of “reconcile” here? (3) L71: What does “reference” mean in the proposed framework? (4) L73: What does an edit refer to? I can understand some of them later when reading the method section, but they remain ambiguous in the introduction section.

  2. Throughout the paper, only two concepts are considered. Do the proposed method and theoretical results only work for two concepts? If so, this would represent a very restrictive scenario. If not, it would be great if the authors could present them in a more general setting.

  3. I do not follow the sentences in L176-179. For evaluation, why do we use R_b at the task level rather than directly using the NLL loss or the probability product like A_{soft}?

  4. Many symbols in this paper are used without definition or introduction. For example, what do D and \alpha_i in L188 mean? And what does the max \emptyset:= 0 in Equation (4) mean?

  5. There are only two digit datasets (MNIST and SVHN) used in the experiments. Standard neuro-symbolic datasets (e.g., the ones used in rsbench) are not considered. [1] Bortolotti, S., Marconato, E., Carraro, T., Morettin, P., Van Krieken, E., Vergari, A., ... & Passerini, A. (2024). A neuro-symbolic benchmark suite for concept quality and reasoning shortcuts. Advances in neural information processing systems, 37, 115861-115905.

Reviewer_wbftOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper extends reasoning-shortcut analysis to incorrect symbolic knowledge. It defines quantities characterizing concept remapping and minimum-edit repair, derives a tight repair radius for full-support addition, and studies the relationship between structural identifiability and neural recovery.

Critical strengths

A concrete, tight structural result. Theorem 3 characterizes the nearest competing zero-loss addition program and establishes a tight worst-case unique-repair radius. This provides an exact answer to a well-defined repair question and is the paper’s strongest substantive contribution.

Critical weaknesses
  1. Limited new understanding beyond a narrow theoretical setting. The strongest result concerns full-support, two-input integer addition with separate slot mappings and unweighted table edits. Beyond this case, the certificates primarily characterize feasible concept mappings and alternative repairs, offering limited insight into which corruptions induce which learned errors or what determines recovery. The general finite-loss bounds are loose (Appendix A.2). There is also an unresolved connection between the discrete zero-loss explanations covered by the certificate (Eq. 5) and the continuous penalized empirical objective used for training (Eq. 6). These limitations leave the broader conceptual contribution insufficiently developed.

  2. Limited explanation of recovery failures. The main recovery gap appears with SmallResNet on SVHN, but the longer-training and optimizer checks concern SlotCNN on the original table. These checks therefore do not establish whether the main gap is robust to training choices. In addition, the cross-table “collapse” measure relies on low per-digit accuracy, which does not distinguish concept merging from reversible relabeling or ordinary classification errors. The interventions provide some evidence that concept quality matters, but only on one table. These limitations leave the broader differences in recovery across architectures and datasets insufficiently explained.

Questions for authors
  1. Beyond the addition theorem, what new understanding of learned concept errors or recovery does the general framework provide?

  2. What do the current experiments establish about the causes of recovery failure, particularly the roles of concept quality and the training setup?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.